This page answers short questions about the copying solution in this book. Each answer is a few lines. Each one links to the section that explains it in full. The method itself is covered by the pushing book's chapter.
Contents#
- Why does copying need two models here, when the pushing book used one?
- Why is the gate's loss weighted by how rare each kind of answer is?
- Why are pushes stored in table coordinates but learned from their glass?
- Why did one fitted copy attempt 36 tables and the others only 9 and 14?
- Why did three times the examples change so little?
- Why does this solution not choose the push direction by arithmetic?
1. Why does copying need two models here, when the pushing book used one?#
In the pushing book every example was a push. Here four answers in five are looks. One model fitted on that mixture learned to look and never pushed. So a small gate now chooses the kind of answer and the glass, and the policy only draws the path. One model decides, one model draws explains the split.
2. Why is the gate's loss weighted by how rare each kind of answer is?#
Without the weights, the first gate never answered "push". That was its safest way to be wrong least often, because pushes are rare. Weighting makes a missed push cost as much as a missed look. Teaching the gate that a rare answer counts gives the change and the gate's scores.
3. Why are pushes stored in table coordinates but learned from their glass?#
Measured from its glass, a push is one movement wherever the glass stands. That makes it easier to learn. But the fine-tuned model reads the same files, and it has no gate to name a glass. So the files keep table coordinates, and this solution converts each push while it trains. A push written from its own glass explains this.
4. Why did one fitted copy attempt 36 tables and the others only 9 and 14?#
That copy gave up quickly. It said 33 of its 36 tables were done while they were still crowded. Stopping early is cheap, so it went far down the queue. Three fits, three very different walks reads that walk.
5. Why did three times the examples change so little?#
The policy had been memorising its tables, and more examples stopped that. Its pushes also came closer to the teacher's direction, and broken glasses halved. But nothing else moved outside the noise, and the gate did not improve. Three times the examples gives the distances.
6. Why does this solution not choose the push direction by arithmetic?#
The blocked direction is its commonest fault, and arithmetic could fix it. But then the solution would be the geometry solution with a learned chooser. It would no longer measure what copying achieves. Why the direction was not handed to arithmetic explains the choice.
← What the repairs bought · A world model, then plan with it — what is different →